The application requires the provisional application sequence number No.60/549 of submission on March 1st, 2004 under 35USC 119 (e), 320 priority, and the disclosure of this application integrally is being hereby incorporated by reference.
Embodiment
Utilization is shown in Fig. 1 according to the digital pre-distortion system of the preferred embodiments of the present invention and the linearisation transmission system of method.
With reference to Fig. 1, the linearisation transmission system comprises digital predistorter (DPD) 100, is used for linearisation and has power amplifier (PA) 110 and traditional digital-analog convertor (DAC) circuit 106 and the transmitter 104 of up-converter circuits 108.Digital input signals x (nT) is coupled with at input 102 places, and is provided to digital predistorter 100.Digital input signals typically can be provided with the plural form with homophase (I) and quadrature (Q) component, as well known in the art, and is imply so here, though only show the individual signals line for convenience of explanation.For example, input signal can be any signal of a plurality of known wide bandwidth signals, such as CDMA that adopts in cellular radio Communication system and WCDMA signal.Digital predistorter 100 receives digital input signals x (nT), and its predistortion is become signal xPD (nT), non-linear with in the compensation transmitter.The predistortion operation masking amplifier of implementing by digital predistorter 100 110 non-linear, and also can randomly proofread and correct any non-linear that other parts by transmitter 104 provide.The analog signal of amplifying typically is provided to the traditional antenna system (not shown) in cellular radio Communication is used at output 112 places.
As shown in Figure 1, digital predistorter 100 comprises three parallel signal paths 114,116 and 118.Path 114 provides linear compensation, and it may only provide time-delay to input signal.Memoryless and memory digital pre-distortion (DPD) circuit 116,118 shows with being separated.As at length discussing the back, memory DPD operation is based on nonlinear multinomial model and memoryless DPD (preferably) is implemented by using look-up table, and look-up table is mapped to input signal amplitude (or power) to the power amplifier gain calibration.Separate and memorylessly allow to use the different structure of gamma correction or different rank with memory DPD circuit.Two kinds of predistortion corrections that provided by memoryless DPD circuit block 116 and memory DPD circuit block 118 are combined at combinational circuit 122 places that can be complex addition circuit, to form the predistortion correction for the combination of input signal.The predistortion correction signal of this combination is added to input signal at main path combinational circuit 120 places that also can be complex addition circuit then, so that the digital signal of predistortion to be provided.The digital signal of this predistortion road 124 along the line is provided to the digital input end of transmitter circuitry 104.
More specifically, the signal x of predistortion
PD(nT) weighted sum of the predefine basic waveform that preferably obtains from input signal x (nT).Basic waveform can be described by one of three kinds: as the linear waveform of the linear function of input signal; The non-linear memoryless waveform that obtains from the instantaneous sample of input signal; Nonlinear function, non-linear waveform as the input sample that in a time interval, obtains based on memory.Relevant with each basic waveform is to be used as complex weighted coefficient.The coefficient weight allows to regulate the amplitude and the phase place of basic waveform before in combination (added together), and is used for making at power amplifier output y
RF(t) distortion minimization in.The selection of basic function and the valuation of coefficient are described in the back.
As shown in Figure 1, correction signal (basic waveform of weighting) is by addition concurrently, to compensate non-linear proterties later in amplifier.The advantage of additive correction signal is that the number of basic waveform can increase by hope, so that improve the linearity of whole system.Just, between the complexity (number of the basic waveform of generation) of performance (correcting value) and predistortion, traded off.
In a preferred embodiment, linear basic waveform equals input signal, if necessary, then by time-delay with in time to quasi-nonlinear basic waveform.Any linear equalization or correction for homophase in the up-conversion block and quadrature error can be compensated in other place in transmission path.Might comprise the part of additional linear basic waveform as DPD; Yet, their any adaptive equalization piece competitions potential and in the transmission path that causes coefficient drift (undesired effect).
As mentioned above, non-linear basic waveform can be classified as memoryless or based on memory.Memoryless basic waveform is characterised in that wide bandwidth, and is that many nonlinear models from input signal obtain.Basic waveform based on memory has narrower bandwidth, and preferably each nonlinear model by the filtering input signal obtains.
In a preferred embodiment, the correction of the memoryless component of non-linearity of power amplifier is to obtain by the look-up table (LUT) that uses the gain error item, and this gain error item is to carry out index by amplitude or the amplitude square of using input signal.The gain error item is to allow the amplitude of accommodation and the phase place complex values as the function of input range.In alternative embodiment, memoryless correction can be that the polynomial expansion formula that the odd-order nonlinear model of input signal and its coefficient weight are the complex scalar items reaches by using its basic waveform.
In a preferred embodiment, the correction of power amplifier memory effect involves the even-order nonlinear model of creating original signal, uses narrow bandwidth filter then, to produce one group of even-order subsignal.One of even-order nonlinear model is the amplitude square of input signal.The even-order subsignal of filtering is the gain error based on memory that input signal is modulated, and proofreaies and correct by the odd-order of wanting thus and produces basic waveform.The coefficient weight is added to each basic waveform, with the amplitude of accommodation and phase place.In a preferred method, there is not the conversion by any way or input signal that provides on the linear path 114 is provided again; Only handle the nonlinear model that obtains from input signal, to be created on the correction signal on path 116 and 118, they are combined at combinational circuit 122 places and input signal.
With reference to Fig. 2, show the self adaptation of utilizing digital pre-distortion coefficient embodiment that generate, linearisation transmission system of the present invention on the figure.The upper path of Fig. 2 is corresponding to the path of Fig. 1, and adopts identical label.The self adaptation embodiment of Fig. 2 adds the feedback path of bottom, and it utilizes the output of input and sampling to provide the DPD parameter of renewal to DPD 100.This is to reach by the coefficient of estimating each basic waveform of weighting.When each iteration, comprise the forward gain of linear signal and distortion from input x (nT) and output y (nT) estimation.
More specifically, as shown in Figure 2, the RF of amplifier 110 output is sampled by sample coupler 200, and the analog rf signal of sampling by traditional analog down converter circuit 202 by down-conversion be demodulated to Simulation with I, Q signal.The up-conversion of analog signal and down-converted are preferably by using identical local oscillator to carry out Phase synchronization for lower frequency changer circuit 202 with up-converter circuit in transmitter 104 in system.The output signal of the sampling of simulation is converted into digital signal by traditional analog-digital converter (ADC) circuit 204.The output of analog-digital converter circuit 204 is on the time with input at time alignment piece 206 aims at, and be normalized into have with in the input at gain block 208 places identical nominal power and sampling rate.Time alignment does not have the constraints of cause and effect, because the input and output sample at first is hunted down, then as a collection of the processing (non real-time).Input signal is provided to forward gain mapping (FGM) circuit block 210, is implemented (schematic diagram of FGM piece is shown in Fig. 4) among the DSP of that it can describe in detail below, as to implement signal processing suitable programming.The product of input reference signal and forward gain provides normalized output y (nT)/G
AveValuation (G wherein
AveBe the nominal gain of stipulating later).Difference between the output of estimating and measure promptly by the error signal of ε (nT) expression, is exported from circuit 212.Coefficient estimator processing block 214 also can be implemented as FGM piece 210 with DSP identical or that separate, is used for obtaining coefficient adjustment, produces least mean-square error.(schematic diagram of coefficient estimator processing block 214 is shown in Fig. 7)
The adaptive digital pre-distortion system of the embodiment of Fig. 2 and method are preferably used alternative manner to estimate and the correction signal that is provided by the DPD parameter that renewal is provided by DPD 100 are provided.Each iteration comprises three steps: (1) is being observed interval IT input and output sample and is being estimated the forward gain error, and (2) are upgraded forward gain error and (3) of accumulation and calculated needed DPD parameter.The sequential of iteration is shown in Fig. 3, and it shows to use observes or the data capture interval 300 and the batch processing in counting period 302, wherein calculates the DPD parameter of upgrading.
Referring again to Fig. 2, in forward gain mapping circuit piece 201, in the observation relevant at interval, obtain the input and output sample with iteration i after, calculate forward gain, so that residual error ε (t) minimizes.Forward gain mapping for iteration i is represented as FGM (i).The forward gain error of accumulation is used for upgrading the predistortion valuation, and this DPD 100 by Fig. 2 implements.
The valuation of the DPD operation of wanting in a preferred embodiment, is the inverse of the forward gain error of accumulation:
(equation 1)
0<α<1 wherein, and []
-1Represent computing reciprocal, describe in the back.When system restrained, forward gain mapping FGM (i) was tending towards one.Therefore, the inverse of the gain error that two important aspects of (equation 1) are the accumulation on forward path and the forward gain error of accumulation is to obtain actual predistortion DPD operation.
The forward gain mapping that is used for iteration i is,
(equation 2) FGM (i)=1+c
0G
Err(i)+G
Mem(i)
G wherein
ErrBe memoryless gain error, G
MemBe based on the gain error of memory, and c
0Be by at G
ErrWith G
MemBetween relevant (describing with respect to coefficient estimator 214 later on) and the plural coefficient weight (near 1) that is affected.The functional block of FGM circuit 210 is shown in Fig. 4 and is implemented (equation 2).Particularly, the input signal at 400 places is provided to three signal paths.402 of first signal paths provide input signal to arrive adder 410 (randomly, having suitable time-delay).The secondary signal path comprises memoryless gain error functional block 404, multiplier 408 and coefficient storage facility 406, second of its enforcement (equation 2) and the result is provided to adder 410.The 3rd signal path comprises the gain error functional block 412 based on memory, the 3rd of its enforcement (equation 2) and output is provided to adder 410.Use the valuation of the output signal of forward gain mapping to be
(equation 3)
G wherein
AveBe preferably by the calculated average gain of following formula:
(equation 4)
Digital pre-distortion gain is that the inversion from the FGM of accumulation obtains:
(equation 5) DPD=1+H
Err+ H
Mem
H wherein
ErrBe the G of accumulation
ErrInverse mapping (in the back describe) and H wherein
MemBe the gain error G based on memory of accumulation
MemInverse mapping (in the back describe).The signal of predistortion is
(equation 6) x
PD(nT)=DPDx (nT).
The functional block of implementing (equation 5) in DPD 100 is shown in Fig. 5.These functional blocks are corresponding to the DPD path 114,116 and 118 of Fig. 1, and use identical label on Fig. 5.
Then, in a preferred embodiment G will be described
ErrAnd H
ErrThe details of determining.In a preferred embodiment, by using input signal amplitude is mapped to the LUT of complex gain error, and implements memoryless correction.The scope of input range is quantized into the storehouse (bin) of fixed number.For each input sample, discern suitable storehouse, and corresponding complex gain error is used for modulation input sample.Usually, the complex gain error changes with input signal amplitude, therefore produces non-linear memoryless waveform.
The index that is used for LUT is
(equation 7) L=round{ ρ [| x (nT) |-q]
Wherein ρ is a scalar item, and q is skew, and round (rounding off) { } is the operator of the immediate integer of identification.The center amplitude of storehouse L is
(equation 8)
Depend on the variation of gain error in the input signal amplitude scope for the number of setting up the needed storehouse of memoryless gain error model.Usually, the highest storehouse should have the maximum input signal of approaching amplitude | x|
MaxThe center amplitude.In this case, the number in storehouse is
(equation 9)
L wherein
MaxIt is the storehouse number.L is determined in the selection of ρ
Max
Complex gain error term is must be by the coefficient of valuation.The input and output sample evidence amplitude of time alignment is grouped into the storehouse.In each storehouse, calculate the average gain error.Gain error for storehouse L is
(equation 10)
W wherein
LBe the window item, it is the function that typically is defined as the input range of following formula:
The window item of regulation makes the storehouse similarly be used in predistortion and valuation in (equation 11).For predistortion, little ρ is preferred, and is less because separate in the storehouse, this so that reduce the quantizing noise relevant again with LUT.Unfortunately, when ρ (storehouse size) hour, the estimation accuracy in (equation 10) is reduced, because have only less measured value to can be used on average.This can cause the big fluctuating in adjacent storehouse.When using (equation 11), exist compromise between quantizing noise in predistorter and the noise sensitivity in the estimator for these two pieces.
In order to remove getting in touch of predistortion quantification and estimation accuracy, LUT is preferably smoothed.Preferable methods is to select a weighting function w for the valuation that the storehouse center distance overlaps
LSuch weighting function is a Hanning window:
(equation 12)
Wherein σ controls the width of Hanning window.Make σ/π compare ρ
-1Much bigger, cause level and smooth LUT mapping.
For storehouse L, the complex gain error of iterations i is represented as G
Err(L, i).Forward gain error for the accumulation of storehouse L is
(equation 13)
Wherein N is an iteration number, 0<α<1, and c
0(i) be the plural coefficient weight that is used for iteration i.For big N numerical value, if iterative process convergence, then G
Err(L is N) near zero.
Should be pointed out that coefficient c
0(i) be that a part as the valuation level of describing in the back obtains, it comprises memoryless and based on the basic waveform of memory.Coefficient c
0(i) be unit 1 for pure memoryless system; For interested input signal format, because memoryless and, with the little deviation that occurs with unit 1 based on the correlation between the correction of memory.
For with signal predistortion, need to be inverted the forward gain error of accumulation.Be inverted preferably and carry out as follows:
(equation 14)
L wherein
WarpedThe center amplitude of expression storehouse L changes.The center amplitude in inverted initial level distortion (warp) storehouse, when gain is low (Gerr<0), compress it and when gaining, expand it to high (Gerr>0): just, by | x|
L, warpedThe new center amplitude of expression becomes
(equation 15) | x|
L, warped=| x|
L| 1+acc_G
Err(L) |.
Using interpolation to come resampling LUT storehouse is easily, and like this, the center amplitude is identical with original storehouse index L.
Calculated gains error, rather than gain have the not advantage of the linear segment of quantizer input signal, and this linear segment is often much bigger compared with needed correction signal.
Then, will describe in a preferred embodiment based on the correction and the G that remember
MemAnd H
MemThe details of determining.Correction based on memory preferably reaches by the model that the nonlinear model that uses filtering is set up based on the gain error of remembering.Even-order nonlinear model for input applies filtering; The signal that finally obtains is transfused to the signal modulation, to obtain basic waveform:
(equation 16) γ
k(nT)=β
k(nT) x (nT)
γ wherein
kBe basic waveform and β
kBe the even-order nonlinear model of filtering, be also referred to as " based on the gain error function of memory ".
Each gain error function based on memory is defined by the exponent number of nonlinear model and the centre frequency of bandpass filtering:
(equation 17)
ω wherein
pBe the centre frequency of filter, m is the pattern exponent number, and g
m(t) be the base band nuclear (baseband kernel) of filter.Base band nuclear and being chosen in of centre frequency describe below.
Frequencies omega
pInfluence the precision of distortion cancellation with the selection of base band nuclear.In a preferred embodiment, the spacing of frequency is fixed for given pattern, but can be different between pattern.Similarly, the base band nuclear that uses in each pattern is identical, but it is different between the numerical value of m usually.
For example, consider the gain error function of second order based on memory:
(equation 18)
For each frequencies omega
pAn independent function β is arranged
2, pQuadravalence gain error function will be:
(equation 19)
Usually, base band nuclear g
2And g
4Be different, the latter is being narrower aspect the time domain internal standard deviation.In addition, for g
4The gain error group of functions compared with g
2Be bigger, because need the more frequency of regulation.This be because, when nonlinear exponent number increased, intermodulation (intermodulation) covered the wideer part of frequency spectrum.As a result, in order to use (equation 17) compensate for slower high-order nonlinear pattern, need more to calculate.
Can define the gain error function based on memory of replacement, use the lower-order gain function to set up the nonlinear model in higher rank.For example, suppose that second-order gain error function group is calculated, shown in (equation 18).Higher-order gain error function can obtain as follows:
(equation 20) alt_ β
M, p(nT)=| x (nT) |
M-2β
2, p(nT).
This repeated use of lower-order result provides the great improvement in the computational efficiency.
This advantage based on the correction of memory is that employed gain error number of functions can be by regulating for the needs of the target that satisfies masking spectrum technical specification (adding surplus).Frequencies omega
pWith base band nuclear g
mSelection determine distortion cancellation performance for given power amplifier and input signal.Wish to select for satisfying the needed number of masking spectrum based on the gain error minimum of a function of remembering.
There is one to trade off and when selecting gain error number of functions and characteristic, take place.For example, excessively specify (over-specifying) gain error number of functions, in the forward modeling of distortion, provide bigger precision; Yet, during the estimation weight coefficient, go wrong.It is difficult more that valuation becomes, and is morbid state (ill-conditioned) because this is separated, or under extreme situation, is not unique.
Consider the nonuniqueness situation.Because basic waveform is not independently, so there is homogeneous solution (homogenous solution).The vector that has the nonzero coefficient that produces zero waveform.This homogeneous coefficient vector can be changed by a scale factor, and does not influence the match of forward gain model, this means, must take additional step to avoid the coefficient drift.Particularly, wish to make the homogeneous coefficient vector minimize for the contribution of the coefficient vector that in the forward gain model, uses.The homogeneous pattern of uncontrollable coefficient vector can cause coefficient to be increased to their maximum, and at this point, it is invalid that the forward gain model will become.
Ill-condition is not too to seriously influence compared with above-mentioned problem, and is that medium excessive appointment owing to the gain error number of functions causes.When separating when being morbid state, estimator makes the forward gain model be fit to random noise (and being not only conclusive crosstalk) degree of freedom that overuses.In long interval, random noise is incoherent with the nonlinear model of input signal.Yet, in short observation at interval, can there be some correlation, this will be included in the coefficient vector of forward gain model (improperly).After forward gain was squeezed, predistortion subsequently made power amplifier properties frequency spectrum degradation, as the noise with generation.
For fear of the problem relevant with nonuniqueness with morbid state, importantly utilize the knowledge of input signal and its spectral characteristic, come for remove to select best gain error function (base band nuclear and frequency) based on the correction of memory.For single carrier wave situation, best base band nuclear relates to pulse shape and is used for generating the time interval of numeric code on the rank of x (nT) and nonlinear model.This can produce uncommon nuclear, such as the Baastians function.For simplicity, preferred embodiment is selected simpler nuclear, such as Hanning function, and the width of regulating nuclear, with the bandwidth on the rank that obtain being applicable to x (nT) and interested nonlinear model.
For the situation of multicarrier input waveform, input x (nT) be on frequency mutually the carrier wave of two or more frequency band limits of skew and value.The bandwidth of each carrier wave makes linear signal separate on frequency typically less than frequency shift (FS).Yet the nonlinear model of composite signal is included in the intermodulation between the carrier wave, causes appearing at the distortion outside original bandwidth of input signal.The bandwidth of intermodulation item and centre frequency can be calculated from the bandwidth and the centre frequency of carrier wave.Be used for generating the frequency difference and the wide intermodulation that is selected to mate between carrier wave of base band nucleus band of the filter of remembering basic waveform.
Usually, when selecting best gain error function based on memory, the multicarrier situation more allows the people be concerned about compared with the single carrier situation.Multicarrier input waveform has wideer bandwidth, and the degradation relevant with the memory effect of power amplifier is the most attractive under wide bandwidth more.
In DPD method of the present invention, by stipulating simple nuclear, such as Hanning window, many taps FIR filter (as shown in Figure 6, being discussed below) is followed in the back, and makes base band nuclear adaptive.The FIR filter comprises a plurality of versions of the nonlinear model of H filtering, each version different amount of being delayed time, and by the coefficient weighting.Adjustment factor weight and the waveform summation to delaying time reach the filtering of wanting.The pattern of handling each time-delay is as the basic waveform of separating, and coefficient is by using the method for describing later by directly valuation.Yet, it seems that from the viewpoint of forward gain model the tandem compound of Hanning window and FIR filter is created adaptive hierarchical topology window, wherein adaptive centre frequency, average delay and the bandwidth that changes window a little of FIR coefficient.This accurate modeling for decisive distortion provides enough flexibilities and simultaneously abundant restraint of liberty degree, to minimize the influence of random noise, in the FIR filter number of coefficient preferably little (for example, show three coefficients on Fig. 6), and preferably half (the N shown in Figure 6 sample) of Hanning window size of spacing of delaying time.Yet, the additional accuracy in the forward model if desired, the number of the spacing of delaying time and coefficient can be changed.
When iteration i, be used for based on the forward gain error model of memory correction be
(equation 21)
C wherein
k(i) be plural coefficient when iteration i, should be pointed out that convenience in order to represent, index n, p replaces with k.The valuation of plural number coefficient is described in the back.The forward gain error coefficient of accumulation is
(equation 22)
0<α<1 wherein.When processing procedure restrains, c
k(i) go to zero.Be approximately the negative value of the forward coefficient of accumulation based on the predistortion of memory:
(equation 23)
Then, show the detailed embodiment of DPD circuit 100 on the figure with reference to Fig. 6.As shown in the figure, the input that provides at 102 places is provided along the first linear signal path 114, and this signal path can be the simple connection to combiner 120, or if necessary, can comprise time-delay, it depends on the specific embodiment of circuit on parallel DPD path.Input signal also is provided to memoryless compensating for path 116, describes with reference to Fig. 1 as above.On memoryless path 116, input is provided to signal magnitude detector circuit 602, and it obtains being provided to look-up table (LUT) 604 corresponding to the signal of the amplitude of input signal with it, look-up table utilizes it to come the LUT clauses and subclauses are carried out index, and this provides above-mentioned memoryless gain error conversion.Alternatively, the signal that is used for LUT is carried out index can be any other appropriate signals relevant with the amplitude of input signal, for example power signal or any function of increasing monotonously with input signal amplitude.The coefficient that is stored in the look-up table 604 will be upgraded periodically by circuit 606, and this circuit 606 receives the coefficient of the renewal on road 608 along the line from coefficient estimator piece 214 in the above self adaptation embodiment that describes with respect to Fig. 2.The output of LUT604 is provided to multiplier 610, and this multiplier receives plural coefficient (above describe with respect to equation 13) from non-volatile storage device 612, so that the output of weighting to be provided.The coefficient that is stored in the memory 612 can be upgraded on road 614 along the line in as the above self adaptation embodiment that describes with respect to Fig. 2 periodically.The output of multiplier 610 is provided to second multiplier 616, and multiplier 616 also receives the input signal on road 618 along the line, with the memoryless basic function component of the more high-order that the digital pre-distortion compensating signal is provided to summing circuit 122.
Input signal from input 102 also is provided to based on the correction DPD path of remembering 118.Input signal is provided to signal power detector 620, and it obtains the signal corresponding to the power of input signal amplitude.Particularly, as shown in the figure, circuit 620 can provide the power signal corresponding to input signal amplitude square, and this power signal output is operated by the hierarchy filter then, and filter is exported or based on the gain error function of remembering, be transfused to the signal modulation.More specifically, the hierarchy filter comprises first fixed coefficient filter 622, provides the baseband filtering operation by using fixing filter kernel; The many taps FIR filter 625 that utilizes adaptive filter coefficient is followed in the back.Fixed coefficient filter 622 can comprise the peaceful filter bank of the Chinese in a preferred embodiment.A specific embodiment of the peaceful filter bank of the Chinese is described below with reference to Fig. 8.In such embodiments, the output of filter bank comprises three real number signals: sine, cosine and DC signal.Therefore, for this embodiment of the peaceful filter of the Chinese that adopts on Fig. 6, the output of filter 622 will comprise three real number signals, though in order to be easy to explanation, only show single line on the figure.Alternatively, the output of filter 622 can be single complex signal or single real number signal.As shown in the figure, the output of filter bank 622 is provided to adaptation coefficient filter 625.First branch 626 of filter 625 receives output road 624 along the line, filter bank 622, and implements the adaptation coefficient filtering operation, and the three rank memory compensation signal outputs on road 656 along the line are provided.Particularly, the signal on circuit 624 is provided to multiplier 632 by first time-delay 628, and the complex filter coefficient that is provided by non-volatile memory cells 634 is provided again this multiplier.In the above self adaptation embodiment that describes with respect to Fig. 2, the plural coefficient in the memory cell 634 is upgraded periodically by coefficient valuation piece along the line 636.Similarly, along the line 624 signal is provided to multiplier 644, and this multiplier receives the complex filter coefficient from non-volatile memory cells 646, and this plural number coefficient road 648 along the line is upgraded periodically.In addition, the signal on road 624 along the line is provided to multiplier 638 via time-delay 630, and this multiplier receives the complex filter coefficient from non-volatile memory cells 640, and this complex filter coefficient road 642 along the line is upgraded periodically.Three multipliers 632,644 and 638 output are provided to adder circuit 650, and this adder circuit is provided to multiplier 652 to the output of filter 626 then, and this multiplier receives the input signal on road 654 along the line.The output road 656 along the line that is transfused to the filtering of signal modulation is provided as three rank memory compensation signals.
Randomly, filter 625 also can comprise additional more high-order memory compensation branch 627.Each such branch preferably includes auto-adaptive fir filter, the more high-order power of the amplitude of this auto-adaptive fir filter receiving inputted signal, more high-order power is filtered and modulated with input signal for these, so that more high-order memory compensation signal to be provided, and for example 5 rank, 7 rank or the like.These of filter 625 more higher order signal branch are represented by circuit block 627 concentrated areas.Particularly, the uniform power of multiplier 658 received signal amplitudes and the output of filter 622, and provide higher rank output signal to arrive many taps FIR filter.This FIR filter comprises: multiplier 664 receives the signal of delaying time from the quilt of time-delay 660 and from the plural coefficient of memory cell 666; Multiplier 676 receives the plural coefficient from memory cell 678; Multiplier 670 receives input signal of delaying time from the quilt of time-delay 662 and the complex filter coefficient that receives from memory cell 672; And adder circuit 682, receive the output of these three multipliers 664,676 and 670.More the output of higher order filter branch is provided to multiplier 684, and this multiplier receives the input signal on road 686 along the line, so that the 3+2m rank memory compensation signal on road 688 along the line to be provided.This more high order FIR filter be adaptive and coefficient storage unit 666,672 and 678 receives the coefficient of along the line 668,674 and 680 renewal respectively.Three rank and more high-order memory compensation signal in adder circuit 690, be combined, the memory compensation signal of combination is provided.The memory compensation signal of this combination is provided to combiner 122.The output of combiner 122 is the memoryless of combination and memory pre-distortion compensated signal, and it is provided to combiner 120, and combined with input signal, and the input signal that predistortion is provided is as output.Those skilled in the art should recognize, the order of fixed coefficient filter and adaptation coefficient filter still can replace equation (20) to integrate with the pattern that is higher than 3 rank and exchanged by user's formula (19).So as used herein, " series connection " of such filter coupling comprises any ordering of such filter.
Then will describe the principle of the operation of coefficient estimator piece 214, the back refers again to Fig. 7 specific embodiment is discussed.This coefficient is preferably by using weighted least mean square (LMS) valuation to calculate.The error signal of sampling is calculated as illustrated in fig. 2:
(equation 24)
Wherein output signal y (nT) is matched with input signal x (nT) by down-conversion, sampling and time alignment on the meaning as the above nominal of describing with respect to Fig. 2.The valuation error can be rewritten as
(equation 25)
Wherein
(equation 26)
Memoryless basic waveform γ
0(nT) be
(equation 27) γ
0(nT)=G
ErrX (nT),
And memory basic waveform γ
k(nT) be
(equation 28) γ
k(nT)=β
k(nT) x (nT)
Or
(equation 29) γ
k(nT)=alt_ β
k(nT) x (nT)
Depend on which kind of form of using the gain error function.Coefficient c based on memory
kThe LMS valuation make | ε (nT) |
2Minimize.
For example consider that wherein three basic waveforms are used for compensating the nonlinear situation based on memory, add memoryless basic waveform γ
0Direct LMS valuation is described below.Measured value is at the time interval [nT-n
0T, nT] the interior accumulation.The coefficient of valuation is
(equation 30)
ε wherein
0, v=[ε
0(nT-n
0T) ... ε
0(nT)]
T, and
(equation 31)
A problem of the direct embodiment of LMS estimator is, this compensation helps having the part of the frequency spectrum of big error power.Unfortunately, this is typically corresponding to the bandwidth across linear signal.Usually, the strictest restriction for spectral emission by government regulation mechanism regulation is beyond the bandwidth that is taken by linear signal.Advantageously valuation is biased to the portions of the spectrum that helps having the tightest emission restriction.
For the valuation of setovering,,, revise error signal and basic waveform such as filter by using linear operation.Because coefficient is a constant, by f
LinearThe linear operator of () expression can be applied to each basic waveform (utilize stack, see Fig. 7) dividually: just
(equation 32)
An example of linear operation is the FIR filter, and it examines h
Est(mT) preferably strengthen as by the portions of the spectrum of the key of correlation standard:
(equation 33)
In (equation 32), also can use such as other such linear operation of iir filter.
Therefore, in order to improve the distortion cancellation in specific portions of the spectrum, with following formula substitution (equation 30):
(equation 34) ε
0, y=[f
Linear{ ε
0(nT-n
oT}...f
Linear{ ε
0(nT) }]
T
With
(equation 35)
When using filtering to stop the linear segment of frequency spectrum, there is the basic waveform of some or all filtering to become the risk of zero (or very near zero).So suggestion is with the valuation regularization, so that stablize valuation: just:
(equation 36)
Wherein R is regular matrix and c
V, defaultIt is default coefficient vector.R and c
V, defaultTypical structure be
(equation 37)
With
(equation 38) c
V, default=[1 00 0]
T.
Usually, the element of matrix R is and at matrix γ
vγ
v TIn elements corresponding compare, usually be little.Yet, in some cases, wish regulation r
00Big numerical value, be unit 1 with the coefficient weighting that forces memoryless basic waveform.
Between the coefficient of memoryless LUT and coefficient, there is interactive possibility during the valuation process, particularly for the input signal of limited bandwidth based on memory.This is that part causes owing to the auto-correlation of input signal shows as the fact that is similar to the time delay expansion relevant with memory effect.If reciprocation occurs with harmful form, it is characterized by fluctuating and memory coefficient increasing absolute value in time at the top of LUT Cang Chu.Below, discuss be used for reducing memoryless with based on the coefficient valuation of remembering between the details of interactive embodiment.
In the last joint, coefficient of utilization c when the valuation of memory coefficient
0, rather than its unit of being set to 1.This is by lowering c to memoryless correction
0Doubly reduce reciprocation.
Another details that is used for reducing interactive embodiment is the input range of regulation less than the LUT of maximum input range.The measured value of the sampling that its input range is right above the I/O in storehouse, top is left in the basket when the valuation of LUT coefficient.Ignore big peak value and prevent that this flip-flop can become problematic at the flip-flop of storehouse, the top LUT of place gain error during gain error is inverted.
The inversion of forward gain error LUT is also restricted.Under most of situations, with the center amplitude of being inverted relevant distortion meeting compression bin.This makes that the inversion gain error of top Cang Chu is undefined.In the present embodiment, regulation is greater than the threshold amplitude of maximum input range.Under this threshold amplitude, the forward gain error of accumulation is set to zero.During the storehouse amplitude that the interpolation level is used for recovering original, the inverted gain error more than the storehouse, top of distortion is filled the value to decay to zero linearly.Rate of decay is determined by amplitude and the difference between the selected threshold amplitude in the distortion in topmost storehouse.As a result, for the big input that surpasses the LUT index range, the correction that is provided by LUT reduces, thereby allows this correction to account for leading by the correction based on memory.This trends towards stoping at LUT and based on the reciprocation between the coefficient of memory.
The gain error LUT of accumulation is preferably smoothed during each iteration before being inverted, help to reduce LUT with based on the reciprocation between the coefficient of remembering.Error by level and smooth introducing reduces in time by iterative process.Steady state solution is not subjected to the gentleest level and smooth influence basically, but the transient state The Characters of coefficient gets better.
With reference to Fig. 7, show the specific embodiment of coefficient estimator piece 214 on the figure with schematic block diagram.As shown in the figure, coefficient estimator piece receives the error signal on the input signal on road 700 along the line and road 702 along the line as input, as above with respect to Fig. 2 and also have (equation 24) to describe.The input signal that road 700 along the line provides is provided to range detector 704, and this range detector obtains the signal corresponding to input signal amplitude, and signal amplitude is used for the memoryless look-up table forward gain of index mapping circuit 706.The output of look-up table 706 mixes with input signal mutually at multiplier 708 places then, provides memoryless basic function as output.This memoryless basic waveform is provided to filter 712, and this filter preferably includes the FIR filter of describing with respect to (equation 33) as above.As noted above, can adopt other suitable linear operator, comprise iir filter.The input signal on road 700 along the line also is provided to the basic waveform maker 710 based on memory, and this maker provides the basic waveform based on memory, and the waveform of N wherein is shown in Fig. 7.The specific embodiment of circuit 710 is shown in the Figure 10 that describes below.Memory basic waveform from the output of waveform maker 710 is provided to each filter block 714-1 to 714-N, these filters preferably also can comprise FIR filter or other the suitable linear operator of describing with respect to (equation 33) as above, comprise iir filter.Filter 716 similarly provides FIR or other suitable linear operation for the error signal that provides at input 702 places.The output of filter is provided to forward gain mapping error coefficient estimator piece 718, and this estimator piece uses above-mentioned lowest mean square to handle the error of determining in the coefficient.The coefficient error amount is provided to then and upgrades piece 720, and this renewal piece coefficient of utilization error provides the coefficient of the correction of renewal, and these coefficients road 216 along the line then is provided to DPD 100, as previously described.
Then, show the preferred embodiment of the peaceful filter bank 622 of the Chinese on the figure with schematic block diagram with reference to Fig. 8.As shown in the figure, filter bank receives input power signal at input 802 places, and this input is split into three signal path, and these signal paths provide three filtering operations that separate.More specifically, first band pass filter 804 and 806 pairs of power signals of second band pass filter provide high and low bandpass filtering operation.The output of first and second band pass filters is used for by utilizing cross-coupled summing circuit 810 and 816, and phase inverter 814 and 90 is spent phase rotation circuits 818 and the cosine and the sinusoidal signal output of generation difference road 812 along the line and 820.Low pass filter 808 so again the delivering power signal the DC component and on circuit 822, provide it as output.Be shown in Fig. 8 by the embodiment of the above peaceful filter of the Chinese common form that finally obtains that provide, the power envelope spectrum, show three common Gauss's output signals of separating that obtain from input power signal on the figure.Should see that additional bandpass filtering can be utilized to provide the additional Gauss's output signal that obtains from input power signal.
Then, show the preferred embodiment based on the basic waveform maker of remembering 710 shown in Figure 7 with a detailed schematic diagram on the figure with reference to Figure 10.As shown in the figure, circuit receives the input signal on road 1000 along the line, and this input signal is provided to signal power detector 1002, and this power detector provides power signal to output to the peaceful filter bank 1004 of the Chinese.The embodiment of Figure 10 is preferably utilized the peaceful filter bank of describing with respect to Fig. 8 such as above of the Chinese, and the output of filter bank 1004 comprises three real number signals, particularly, sine, cosine and DC signal as shown in the figure, that road 1006,1008 and 1010 along the line provides.The output of the peaceful filter bank of the Chinese is provided to three FIR filters that separate, and these filters provide filtering operation to input signal, and provides as shown in the figure each basic waveform as output.Because the operation of each filter is identical, so will only describe first filter.As shown in the figure, filter comprises first filter branch with time-delay 1012, this time-delay 1012 is provided to multiplier 1014 to its output, this multiplier is receiving inputted signal also, the first memory basic waveform output is provided then, second filter branch is provided to multiplier 1016 to the signal on the circuit 1006, this multiplier is receiving inputted signal also, the second memory basic waveform output is provided then, the 3rd filter branch comprises time-delay 1018, and this time-delay 1018 is provided to multiplier 1020 to its output, and this multiplier is receiving inputted signal also, provide the 3rd memory basic waveform output then, as shown in the figure.Though show 9 memory basic functions on Figure 10, can adopt the basic function of interpolation or less basic function.
In alternative embodiment, above-mentioned adaptation coefficient estimator function can be utilized pre-distortion coefficients tabulation and the relevant list management program of implementing in the DSP of suitably programming, functional together with coefficient estimator piece 214.The best coefficient that is used for digital pre-distortion changes with average input range (or power), temperature, pattern of the input (now using the number and the frequency of carrier wave) and other measurable input or environment parameter.In above-mentioned method, the change of optimum coefficient is followed the tracks of by the adaptive characteristic of system.Yet, might be relevant the coefficient of success in the past with input and environment parameter by the tabulation that forms successful coefficient vector in the past, wherein each vector is assigned with an attribute vector.Attribute vector is used as multi-dimensional indexing, comprises input range, pattern of the input, temperature and other measurable input or environment parameter.When input or the change of environment parameter enough make the predistortion performance degradation greatly, the coefficient vector of retrieving novel from the tabulation of attribute vector with the attribute that approaches most current measurement.Iterative process uses this new coefficient vector as initial starting point.This tabulating method is at the situation that is used for feedforward compensation, on January 21st, 2004 submitted to, U.S. Patent Application Serial Number No.10/761, in 788 and be used for linearizing, the U.S. Patent Application Serial Number No.10/889 of self-adapted pre-distortion, describe in 636, the disclosure of these patent applications integrally is being hereby incorporated by reference.Particularly, can directly adopt at U.S. Patent Application Serial Number No.10/889, the function of describing in 636 comprises forming and keeping the pre-distortion coefficients tabulation.
When forming the pre-distortion coefficients tabulation, must select is all coefficients that storage comprises memoryless LUT clauses and subclauses, and still a storage (comprises coefficient c based on the coefficient of memory
0).When the regularization in (equation 36) was applied energetically, memoryless LUT clauses and subclauses trended towards changing very little; In this case, only need the coefficient of storage based on memory.
When using the pre-distortion coefficients tabulation, must be given for the similarity of two attribute vectors of judgement and the homophylic distance measure of two coefficient vectors.Attributive distance is estimated at above-mentioned U.S. Patent Application Serial Number No.10/889, and (and in U.S. Patent application 10/761,788) described in 636.The coefficient distance measure is deleted in one of process identification and deletion redundancy unit (attribute and coefficient vector to) two and is used.The coefficient distance measure can comprise memory LUT and memory coefficient; Yet for simplicity, the suggestion distance is only based on memory coefficient difference.
Also must create and determine that quality that digital pre-distortion is proofreaied and correct and regulation determine that when correction is the measure of successful threshold value.Correction mass relates to the relevant masking spectrum by government bodies' regulation ideally.Successful correction transmission has the requirement of sheltering of enough surpluses.Yet, function (the E{| ε that easier regulation correction mass is remaining square error
0|
2, wherein E{} represents desired value) or the error (E{|f of remaining square filtering
Linear{ ε
0|
2), they are calculated during the valuation process.These residual value typically by input signal power (or other input measurement value) by normalization.Also can be used for judging the quality of correction for the forward gain error coefficient (be not accumulation forward gain error) of the iteration of regulation, because they go to zero when iterative process converges to its optimum value.L
2Or L
InfNorm can be applied to the forward gain error coefficient and with the threshold of selecting, to determine to proofread and correct whether success.
Therefore the present invention provides digital pre-distortion system and method, has a plurality of characteristics and advantage, comprising: use gain error correction, so that the influence that LUT quantizes minimizes; Make up memoryless LUT proofread and correct with based on polynomial correction based on memory; Use hierarchy filtering by creating the sef-adapting filter structure that can compensate decisive distortion, so that improve correction based on memory, and do not introduce attempt improperly to random signal handle the falseness of modeling the degree of freedom (this so will increase noise floor, so be undesired); More the embodiment of high-order memory compensation is remembered the computational efficiency that the result strengthens embodiment by reusing lower-order; Use least-square methods to come estimation coefficient, it is weighted so that provide bigger correction (use the frequency spectrum weighting, usually, its reflection is by the masking spectrum restriction of government bodies' regulation) in the different part of frequency spectrum; And use the level and smooth and restriction of the gain error in LUT is inverted of LUT and avoid undesired reciprocation between the memory module of LUT and reaction.
Though described specific embodiment and embodiment details, and do not meant that in itself and limit,, it will be appreciated that as those skilled in the art because many variations and modification can be provided.